Mismatched image identification using histogram of loop closure error for feature-based optical mapping

Image registration is one of the most fundamental steps in optical mapping from mobile platforms. Lately, image registration is performed by detecting salient points in two images and matching their descriptors. Robust methods [such as Random Sample Consensus (RANSAC)] are employed to eliminate outliers and compute the geometric transformation between the coordinate frames of images, typically a homography when the images contain views of a flat area. However, the image registration pipeline can sometimes provide a sufficient number of wrong inliers within the error bounds even when images do not overlap at all. Such mismatches occur especially when the scene has repetitive texture and shows structural similarity. Such pairs prevent the trajectory (thus, a mosaic) from being estimated accurately. In this paper, we propose to utilize closed-loop constraints for identifying mismatches. Cycles appear when the camera revisits an area that was imaged before, which is a common practice especially for mapping purposes. The proposed method exploits the fact that images forming a cycle should have an identity mapping when all the homographies between images in the cycle are multiplied. Our proposal obtains error statistics for each matched image pair extracting several cycle bases. Then, by using a previously trained classifier, it identifies image pairs by comparing error histograms. We present experimental results with different image sequences.